Summary: MiniMax released M2.5 and M2.5-Lightning on February 12, 2026, posting open weights the same day. M2.5 scores 80.2% on SWE-Bench Verified, averages 22.8 minutes per task, and costs about a tenth of Claude Opus 4.6, while Lightning doubles speed and price. The 229B-parameter weights use a modified MIT license and support SGLang, vLLM, Transformers, and KTransformers deployment.
On February 12, 2026, MiniMax put out two models, M2.5 and M2.5-Lightning, and posted the weights to Hugging Face the same day. The official positioning compresses into one sentence: capability near the frontier, at a price that is a fraction of it. The model posts 80.2% on SWE-Bench Verified and clears a task in 22.8 minutes on average — level with Claude Opus 4.6’s 22.9 minutes — for roughly a tenth of the spend.
The launch lands in the middle of a crowded stretch of Chinese model releases. Zhipu had open-sourced GLM-5 one day earlier (see our earlier coverage), and VentureBeat framed its M2.5 story around near-state-of-the-art performance at a twentieth of Claude Opus 4.6’s cost. Cheap plus open is no longer one vendor’s playbook; it is the baseline the whole market ships with.
One Capability, Two Speeds
There is nothing to choose between M2.5 and M2.5-Lightning on capability; the split is purely speed and price. The standard M2.5 runs at 50 tokens per second, at $0.15 per million input tokens and $1.20 per million output tokens. Lightning doubles that to 100 tokens per second — twice most frontier models, by MiniMax’s claim — at $0.30 input and $2.40 output, with caching available on both variants. The announcement translates the rates into costs you can feel: an hour of continuous Lightning runs $1, M2.5 just $0.30, and four always-on instances kept up for a full year total $10,000. The model card calls this “intelligence too cheap to meter.” For scale, a single SWE-Bench Verified task consumes about 3.52 million tokens on average — a workload that would sting at frontier-model pricing and registers as routine here.
Coding Benchmarks Beyond the Headline Number
Past the 80.2% on SWE-Bench Verified, the persuasive part is the cross-harness record: on Droid, M2.5 scores 79.7 against Opus 4.6’s 78.9; on OpenCode it is 76.1 to 75.9. Any single benchmark can be gamed; holding steady across environments — including harnesses the model was not tuned for — is what agent engineers actually pay for. The rest of the sheet: 51.3% on Multi-SWE-Bench, 76.3% on BrowseComp with context management, roughly 20% fewer search and tool-calling rounds than M2.1, and a VIBE-Pro result MiniMax describes as on par with Opus 4.5. For office work, the reported figure is a 59.0% average win rate on MiniMax’s GDPval-MM evaluation, tied to the Office Skills for Word, PowerPoint, and Excel built into MiniMax Agent, plus more than ten thousand user-built Experts.
The Recipe: Forge, CISPO, and Agentic RL at Scale
Underneath, the training recipe mirrors the playbook this wave of Chinese models shares: agentic reinforcement learning at scale. MiniMax ran RL across more than 200,000 real-world coding environments spanning 10-plus languages and full-stack work on Web, Android, iOS, and Windows. Two homegrown pieces carry the load: Forge, an agent-native RL framework that decouples agents from the training-inference engine and pairs tree-structured sample merging with async scheduling for roughly a 40x training speedup; and CISPO, an algorithm that keeps large-scale MoE training stable. Reward design mixes process rewards with trajectory-level task-completion-time optimization — the model is forced to trade off being smart against being fast, which is where that 22.8-minute average comes from.
Weights on Hugging Face, and What Self-Hosting Takes
The weights are on Hugging Face under a modified MIT license: 229B parameters, stored mostly in FP8 quantization, roughly 230GB to pull down. Deployment runs through the usual paths — SGLang, vLLM, Transformers, and KTransformers. Put MiniMax’s internal numbers next to that — M2.5 now autonomously handles 30% of tasks inside the company and accounts for 80% of newly committed code — and the bar for running your own “primary coding model” has, for the first time, fallen to a multi-GPU server plus an open weights download.
Where This Leaves Developers
Three observations. First, the marginal cost of coding agents is collapsing: output pricing sits at a tenth to a twentieth of Opus-class models, which turns a long-running agent from a cost-control problem into overhead you stop measuring. Second, open weights reset procurement: a team that can self-host 229B now carries an anchor into every API negotiation. Third, speed itself is now a spec: Lightning’s 100 tokens per second brings interactive agents close to the feel of local tooling — at once a new constraint and a differentiator for IDE and terminal agent UX.